DevOps
The Prompt and the Model That Could Not Be Separated
The rollback reverted both changes at once, so the fix for the prompt also undid the model upgrade, and the incident lasted until someone split the two and…
Rohit Bhadani
I write about why software breaks and how people figure it out. Founder of Krova Cloud, building isolated microVMs on our own hypervisor. Years of Shopify theme work and shared hosting before that. If I have not reproduced it, I do not publish it.
Tooling
The Span Per Tool Call That Made One Agent Step Debuggable
The trace had one span for the whole run, which explained nothing, and one span per tool call turned an opaque minute into a readable sequence with a single…
Code Review
Reviewing the Loop That Fired Ten Thousand Model Calls at Once
The batch loop awaited every item concurrently with no semaphore, so the first burst rate limited itself, and the retry storm that followed turned a…
Opinion
Your AI Feature Needs a Fallback That Is Not an Apology
When the model is unavailable the feature should still do something useful, and an apology message is not a fallback, it is an outage with better copy, and…
Deep Dive
Why LLM Latency Is Bimodal and the Average Hides It
The response times split into a fast cluster and a slow cluster, and the mean landed in the empty space between them, describing a latency that no request…
Bug Hunt
The Embedding Swap That Silently Broke Semantic Search
A newer embedding model went live for queries while the index still held the old vectors, and the two spaces were unrelated, so similarity became noise and…
Error Autopsy
The Token Budget Kill Switch That Never Fired
The guard was armed, the alert was wired, and the switch never fired, because it compared a token count against a dollar limit, and the units never matched…
Ecommerce
The AI Product Copy That Promised a Warranty Nobody Offered
Generated descriptions filled the gaps in the spec with plausible terms, and the plausible terms became a promise on three hundred product pages, and the…
Performance
The Context Stuffing That Inflated Time to First Token
Every retrieved document went into every prompt because retrieval was cheap, and the prefill cost of the context nobody read became the latency users waited…
DevOps
The Model Upgrade That Changed the Output Schema Nobody Pinned
The model version floated, the provider shipped an update, and the JSON the parser expected gained a wrapper, so every request parsed to empty and nothing…
Tooling
The Golden Dataset That Caught a Quiet Model Regression
A frozen set of inputs and expected behaviours turned a model upgrade from a hope into a measurement, and the measurement showed a drop nobody had asked…
Code Review
Reviewing the Off-by-One the Model Put in the Pagination Loop
The loop fetched every page except the last, and the omission was invisible because the page count was right and the total was right, and the only wrong…